The List / Agents and Multi-Agent Systems / paper
Can Large Language Models Mine Interpretable Financial Factors More Effectively? A Neural-Symbolic Factor Mining Agent Model
Abstract
Finding interpretable factors for stock returns is the most vital issue in the empirical asset pricing domain. As data-driven methods, existing factor mining models can be categorized into symbol-based and neural-based models. Symbol-based models are interpretable but inefficient, while neural-based approaches are efficient but lack interpretability. Hence, mining interpretable factors effectively presents a significant challenge. Inspired by the success of Large Language Models (LLMs) in various tasks, we propose a FActor Mining Agent (FAMA) model that enables LLMs to integrate the strengths of both neural and symbolic models for factor mining. In this paper, FAMA consists of two main components: Cross-Sample Selection (CSS) and Chain-of-Experience (CoE). CSS addresses the homogeneity challenges in LLMs during factor mining by assimilating diverse factors as in-context samples, whereas CoE enables LLMs to leverage past successful mining experiences, expe-diting the mining of effective factors. Experimental evaluations on real-world stock market data demonstrate the effectiveness of our approach by surpassing the SOTA RankIC by 0.006 and RankICIR by 0.105 in predicting S&P 500 returns
In one sentence
Summary
The paper addresses the challenge of mining financially interpretable factors for stock return prediction, a critical task in empirical asset pricing. Existing methods are divided into symbolic models, which offer interpretability but suffer from inefficient search spaces, and neural models, which are efficient but lack transparency. The authors propose FAMA (FActor Mining Agent), an agent-based framework that leverages Large Language Models (LLMs) to bridge this gap. FAMA treats the LLM as a neuro-symbolic bridge, using in-context learning to generate symbolic factor expressions that retain financial interpretability while benefiting from the generative power of neural networks. The core problem is mitigating the homogeneity and inefficiency of direct LLM factor generation by structuring the prompt with diverse examples and past successful mining paths. The FAMA model consists of two key components: Cross-Sample Selection (CSS) and Chain-of-Experience (CoE). CSS employs K-Means clustering on factor exposures to select low-correlation factors as in-context examples, ensuring diversity in the generated factors and reducing homogeneity. CoE maintains a dynamic chain of successful factor mining paths, allowing the LLM to learn from past experiences. The system iteratively mines factors using text-davinci-002 on S&P 500 stock data from 2015 to 2022. Experiments compare FAMA against symbolic baselines like Alpha101 and Genetic Programming, and neural baselines like FactorVAE and DTransformer. Evaluation metrics include RankIC and RankICIR for predictive accuracy, and annualized return and Sharpe ratio for investment simulation. Results demonstrate that FAMA outperforms state-of-the-art models, achieving a RankIC improvement of 0.006 and RankICIR of 0.105 over the best neural baseline. In investment simulations, FAMA achieved an annualized return of 38.4% and a Sharpe ratio of 667.2%, significantly surpassing the S&P 500 benchmark and other models. The ablation studies confirm that both CSS and CoE contribute to performance, with CoE iterations enhancing results up to a point. However, the paper notes limitations regarding LLM hallucinations in the financial domain, which can introduce noise into the factor mining process. The study highlights the potential of agentic workflows for alpha mining but cautions about the need for robust error handling in production environments.
Research problem
The paper addresses the difficulty of mining financial factors that are both highly predictive and financially interpretable. Traditional symbolic methods are interpretable but inefficient due to large search spaces, while neural methods are efficient but produce opaque features. The specific problem is how to leverage LLMs to generate interpretable symbolic factors effectively, overcoming the homogeneity and inefficiency of direct in-context learning approaches.
Core contributions
- Proposes FAMA, the first agent-based model using LLMs as a neuro-symbolic bridge for factor mining.
- Introduces Cross-Sample Selection (CSS) using K-Means clustering to ensure diversity in in-context examples.
- Develops Chain-of-Experience (CoE) to incorporate past successful mining paths into the LLM prompt.
- Demonstrates superior predictive performance (RankIC/RankICIR) compared to symbolic and neural baselines.
- Shows significant improvement in investment simulation metrics (AR and Sharpe Ratio) on S&P 500 data.
Method
- Factor clustering using K-Means on z-score normalized factor exposures.
- Cross-Sample Selection (CSS) picks low-correlation factors from different clusters as in-context examples.
- Chain-of-Experience (CoE) tracks successful factor mining paths and updates them iteratively.
- LLM prompt combines CSS examples and CoE paths to generate new symbolic factor expressions.
- Iterative mining loop: generate factor, evaluate RankIC, update CoE if performance improves.
- Portfolio simulation using top 20% factor value stocks with weights based on past RankIC.
Results & findings
- FAMA(CI-7) achieved RankIC of 0.054 and RankICIR of 0.485, outperforming FactorVAE (0.048/0.379).
- FAMA surpassed the S&P 500 benchmark in investment simulation with 38.4% annualized return vs 26.3%.
- Sharpe ratio for FAMA was 667.2%, significantly higher than FactorVAE's 132.2% and S&P 500's 209.3%.
- CSS and CoE ablation shows both components are necessary for optimal performance.
- Removing initial nodes from CoE chains improved performance, suggesting low-performing factors hinder LLMs.
- Increasing CSS samples beyond 3 did not improve performance, indicating a threshold for diversity.
- FAMA used only 10% of the training data for factor generation, showing data efficiency.
Data & experiments
- Dataset: S&P 500 stock data (close, open, high, low, volume) from 2015 to 2022.
- Initial factors: 38 factors from Alpha101 used as the starting set.
- Baselines: Alpha101, Genetic Programming (GP), ALSTM, DTransformer, FactorVAE.
- Metrics: RankIC, RankICIR for prediction; Annualized Return, Volatility, Sharpe Ratio for investment.
- LLM: text-davinci-002 with temperature 0 and max_tokens 1500.
- Ablation studies on CSS sample size and CoE iteration depth.
Limitations & caveats
- LLM hallucinations can introduce noise into the factor mining process.
- Performance may vary across different market regimes or asset classes.
- Dependence on text-davinci-002, which is deprecated; results may differ with newer models.
- Simulation assumes no transaction costs or slippage.
- Risk-free rate set to zero for simplicity, which may not reflect real-world conditions.
- Limited to S&P 500 stocks; generalizability to other markets is not tested.
Curator notes
Discovery status: institutional_trading_round2_promoted. Source links: From Natural Language to Executable Option Strategies via Large Language Models